Skill packs
The specialists are horizontal by default. Skill packs are how you make them vertical — a restaurant pack, a real-estate pack, a SaaS pack, a newsletter pack.
TL;DR
A skill pack is a bundle of vertical expertise — prompt fragments, templates, small tool belts — that attaches to the right specialists only when relevant. The Coding Specialist gets your SaaS-growth templates; the Legal Specialist gets your real-estate compliance checklist. The Evaluator, deliberately, gets neither.
Why skill packs exist
The core specialists know how to do their job. They do not know your industry. A Content Specialist can write copy, but the phrasing and regulatory caveats for a mortgage broker are different from those for a yoga studio. Baking every vertical into every specialist's system prompt would make them slower, blander, and less safe.
Skill packs solve this by making vertical knowledge optional and composable. A consultant installs the "consulting-operations" pack; a restaurant installs "restaurant-ops." The core specialists stay lean; the vertical knowledge shows up only where it matters.
What a skill pack contains
- A prompt-fragment addendum. Role-specific know-how the specialist should keep in mind — how a real-estate disclosure reads in California, what a SaaS cold email should not say.
- A tool belt. Optional — a small set of MCP servers or APIs. The restaurant pack might include a menu-pricing tool; the podcaster pack might include a transcript tool.
- Templates and checklists. Document formats, starter assets, compliance checklists the specialist can copy and adapt.
- A target list. Which specialists this pack attaches to. Most packs target two or three, not all of them.
When skill packs activate
Skill packs are not loaded into the base specialist system
prompts. That would make every conversation longer and
slower. They activate at delegation time: when the CEO calls
delegate_to_content for a task that the active
skill packs target, the relevant pack fragments are spliced
into the sub-session's system prompt and the pack's tools
are added to the allowed-tools list. Different tasks may pull
different combinations.
This also means you can have multiple skill packs installed at once. A consultant who also runs a podcast installs the consulting pack AND the podcasting pack; the Content Specialist picks up whichever is relevant to the specific task, or composes them when both apply.
The Evaluator does not get skill packs
One specialist is deliberately excluded: the Evaluator. The
Evaluator's job is to be an independent skeptic — to
second-guess whatever the other specialists produced. If the
same skill pack that shaped the Content Specialist also shaped
the Evaluator, both would share the same blind spots. The
SKILL_PACK_TARGET_SPECIALISTS list in
apps/engine/src/skills/loader.ts explicitly omits
the Evaluator for this reason.
How skill packs differ from connections
A connection is an auth-and-credentials bridge to an external tool (your Gmail, your GitHub, your Stripe). A skill pack is content — it tells a specialist what to do once it has a connection. The two work together: the real-estate pack explains how to draft a listing, and the MLS connection is how it gets published. Packs without connections are informational; connections without packs are generic.
How it works in Black Box
Skill packs live under apps/engine/src/skills/.
Each pack is a small folder with a manifest (metadata, target
specialists), one or more prompt fragments, optional tool
configs, and optional template files. The loader
(skills/loader.ts) reads the manifest, matches it
against the current delegation, and wires the fragments +
tools into the specialist's query() call at
runtime.
From the owner's side, packs are managed in the Skills panel. You can install, deactivate, and update packs. The CEO will also proactively suggest packs based on company memory — if you run a newsletter, it will suggest the newsletter pack; if you talk about Shopify, it will suggest the e-commerce pack. You can always say no.
Authoring skill packs
Today, Web4Guru ships packs on our own schedule, and enterprise customers contract custom packs. Self-serve authoring is on the roadmap: a YAML + prompt-fragment format that anyone can write in a text editor and submit for review. Once the registry is live, the community will be the long-run source of specialization — the same pattern that made VSCode extensions dominate.
Frequently asked
- What is in a pack?
- Prompt fragments, a small tool belt, templates, and a target-specialist list.
- Are packs always on?
- No. They activate at delegation time and only on their target specialists.
- Does the Evaluator get packs?
- No. Deliberately — an independent skeptic cannot share the reviewed party's blind spots.
- Can I author one?
- Enterprise today; self-serve on the roadmap.
Related concepts
- The 18 specialists — who skill packs attach to.
- Playbooks — scripted sequences that often pull in specific packs.
- Connections — the tool-auth layer that complements packs.
- Feature page: skill packs.
- Glossary: AI skill pack.
Next: playbooks
Pre-wired sequences for the outcomes Black Box ships most often. Less improvisation, more reliability.